litert-community/SmolLM3-3B

Chat template: accept the 0.18 content-parts form (string form unchanged); weights, tokenizer and executor metadata byte-identical

Card: measured device / browser results (edge-compat, 2026-09)

manifest: Android measured rows (Galaxy S26 GPU + same-device CPU control; Pixel 8a where run) and an Android recommendation where the evidence supports one

Card: definition line (LiteRT, 2026-09)

manifest: Galaxy S26 GPU rows from the S4 gate sweep (Android recommendation where none existed)

Card: base_model_relation: quantized (list under the base model's Quantizations)

Add measured Raspberry Pi 5 (CPU) section

Manifest: regenerate after the thought-channel re-pack (sha256 + capabilities.thinking)

Card: note the thought-channel declaration (metadata only, weights unchanged)

Declare the thought channel so the reasoning is separable and the thinking budget applies (metadata only; weights byte-identical)

Declare the thought channel so the reasoning is separable and the thinking budget applies (metadata only; weights byte-identical)

Card: note the tokenizer-section fix (weights unchanged)

Replace the tokenizer section with the upstream tokenizer.json (weights, graph and metadata unchanged). The previous SentencePiece conversion of the BPE tokenizer encoded standalone accented/special characters to the wrong ids, turned characters without a vocabulary entry (emoji, some accented capitals) into the end-of-text/end-of-turn token, and could not match the token it had reused as UNK; prompts now tokenize identically to the upstream tokenizer (verified on the LiteRT-LM runtime, see the card note). litertlm_manifest.json regenerated in the same commit.

Card: note the default-system-prompt fix (weights unchanged)

Restore the vendor's default system prompt in the chat template (weights unchanged). The converter's template probe never renders the block the upstream template emits when the caller sends no system message, so the bundle silently dropped it; with no system message the prompt now renders byte-identical to the upstream chat template's output (see the card note). litertlm_manifest.json regenerated in the same commit.

Restore the vendor's default system prompt in the chat template (weights unchanged). The converter's template probe never renders the block the upstream template emits when the caller sends no system message, so the bundle silently dropped it; with no system message the prompt now renders byte-identical to the upstream chat template's output (see the card note). litertlm_manifest.json regenerated in the same commit.

Card: correct the start_token note — measured numbers do not carry over

Card: note the start_token fix (weights unchanged)

Regenerate litertlm_manifest.json for the new bundle sha256

Drop the metadata start_token (weights unchanged)

Add measured Galaxy S26 GPU backend section

Add litertlm_manifest.json — machine-readable deployment manifest (variant selection, backend recommendations, measured performance)

Add measured Pixel 8a OpenCL delegation (runtime capability; Gallery guidance unchanged)

Card: fill the iPhone prefill/TTFT from the run log (they were recorded; the previous note was wrong)

Card: measured Performance table (M4 Max CPU/GPU + on-device where measured), accuracy note

Add model (#1)

Document Gallery 1.0.16 direct Hugging Face import + desktop LiteRT-LM CLI (serve/run)

docs: add Training data & PII section (submission-guideline compliance)

Upload folder using huggingface_hub

initial commit